Accelerating Materials Science with Big Data and Machine Learning: On-Demand Oral Presentations
Program Organizers: Huan Tran, Georgia Institute Of Technology; Muratahan Aykol, Toyota Research Institute
Friday 8:00 AM
October 22, 2021
Room: On-Demand Room 2
Location: MS&T On Demand
Invited
Bridging the Gap between Literature Data Extraction and Domain Specific Materials Informatics: Elsa Olivetti1; 1Massachusetts Institute of Technology
Data has become a fundamental ingredient for accelerating and optimizing materials design and synthesis. Advances in applying natural language processing (NLP) to material science text has greatly increased the size and acquisition speed of materials science data from the published literature. This presentation will describe work to extract information from peer reviewed academic literature across a range of materials. Applying NLP pipelines to these types of materials science systems can be challenging due to the general schema and the noisiness of automatically extraction data. I will present data engineering techniques and discuss an optimal balance between automatic and manual data extraction.
Invited
There is No Time for Science as Usual: Alan Aspuru-Guzik1; 1University of Toronto
The world is facing several time-sensitive issues ranging from climate change to the rapid degradation of our climate, as well as the emergence of new diseases like COVID-19. We need to rethink the way we do science and think of it as a workflow that could be optimized. Where are the pain points that can be solved with automation, artificial intelligence, or better human practices? My group has been thinking about this question with an application to the design of organic optoelectronic materials. In this talk, I will discuss the progress in developing materials acceleration platforms, or self-driving labs for this purpose.
Invited
Designing Alloys with Process-mapping AI Pre-trained on Empirical Knowledge: Vyacheslav Romanov1; 1National Energy Technology Laboratory
Accelerated materials design should match the recent trends in the product development cycles. Materials data analytics can be used to significantly shorten development time of specialized alloys needed for next generation energy applications. However, it faces a challenge of scarce data available for training ML models. Incorporation of the domain knowledge into deep-learning graph structure via fuzzy pre-training and causal process imitation presents a viable approach to developing accurate data-driven models and reliable alloy design tools, with limited datasets. Artificial Intelligence (AI) was used in this study to incorporate such knowledge in the domain-specific computational tool, pyroMind. The tool provides not only novel design ideas but also their interpretation via physics and engineering concepts.
Invited
Accelerating Discovery in Computational Materials Science Using CAMD: Joseph Montoya1; 1Toyota Research Institute
Artificial intelligence and machine learning are enabling automation of decision-making in various scientific domains, but still face a number of fundamental obstacles in materials science. We provide an overview of one such platform, Computational Autonomy for Materials Discovery (CAMD), designed to help materials scientists simulate and design their discovery processes using machine learning tools. CAMD has specifically been engineered to maximize the likelihood that sequential iterations of an experimental or simulation-based workflow will produce materials data with target properties. To date, CAMD's primary application is in the prediction of new, phase-stable crystal structures from structural prototypes in various chemical spaces. In addition, we have begun designing multi-fidelity sequential learning agents using data streams from experiment and theory. We review these capabilities with a view towards the future of AI-assisted tools for materials discovery.
Scalable Gaussian Processes for Predicting the Optical, Physical, Thermal, and Mechanical Properties of Inorganic Glasses Using Compositions for Large Datasets: Suresh Bishnoi1; Ravinder Ravinder1; Hargun Singh Grover1; Hariprasad Kodamana1; N. M. Anoop Krishnan1; 1Indian Institute of Technology, Delhi
Gaussian process regression (GPR) is an extremely useful technique to predict composition–property relationships in glasses. The GPR’s main advantage over other machine learning methods is its inherent ability to provide the standard deviation of the predictions. However, the method remains restricted to small datasets due to cubic time complexity associated with it. So herein, using a scalable GPR algorithm, namely, kernel interpolation for scalable structured Gaussian processes (KISS-GP) along with massively scalable GP (MSGP), we develop composition–property models for inorganic glasses. The models are based on a large dataset with more than 100000 glass compositions, 37 components, and nine crucial properties: density, Young’s, shear, bulk moduli, thermal expansion coefficient, Vickers’ hardness, refractive index, glass transition temperature, and liquidus temperature. We show that the models developed here are superior to the state-of-the-art machine learning models. We also demonstrate that the GPR models can reasonably capture the underlying composition-dependent physics.
Deep Learning-enabled Prediction of Mechanical Properties of Metallic Microlattice Structures Using Uniaxial Compression Videos: Akanksh Shetty1; Chunshan Hu1; Mohammad Sadeq Saleh1; Jack Beuth1; Rahul Panat1; Amir Farimani1; 1Carnegie Mellon University
Mechanical properties of cellular materials are important in fields such as lightweight-materials, bone-implants, and energy storage devices. Obtaining force-displacement curves from experiments, however, is costly, and time consuming. FEA is computationally intensive and theoretical models cannot fully capture various geometries of cellular materials. In this paper, metallic microlattices were fabricated by Aerosol Jet printing and subjected to uniaxial compression. High-resolution videos of the compression tests, along with measured force-displacement curves were used to train dataset for a Convolutional-Neural-Network-Long Short-Term-Memory-Network(CNN-LSTM) model. Force-displacement curve was predicted based on compression videos of untrained samples and was compared with experimental data. To further improve the performance, physics-based features were extracted from the videos and used for the training of LSTMs. Excellent prediction capability is demonstrated with average-Intersection-Over-Union score of >0.95 for train test split of 0.1. This study demonstrates that deep learning can be used to accurately predict the mechanical behavior of cellular materials.
Molecular Dynamics Simulation Using Lagrangian Neural Networks: Ravinder Bhattoo1; N. M. Anoop Krishnan1; 1Indian Institute of Technology Delhi
The accurate interatomic potential energy functions (PEF) are critical for valid molecular dynamics simulations. The interatomic PEF is developed by parameterizing a functional form using experimental data or DFT (Density Functional Theory) simulation data. Therefore, estimating a functional form is critical in determining the interatomic PEF. Herein, we use a neural network (NN) to define the Lagrangian function of an atomic system as NN is well known as a universal function approximator. Further, we use the Euler-Lagrange equation to determine the acceleration of the atomic system. Finally, we train the NN against the simulation trajectories of unary and binary Lennard Jones (LJ) systems created from traditional molecular dynamics (MD) simulations. The trained NN is then used to do MD simulations and demonstrate the energy conservation.
Multi-target Prediction of Concrete Engineering Properties Based on a Single Deep Learning Model: Yu Song1; Gaurav Sant1; Mathieu Bauchy1; 1University of California, Los Angeles
Despite the recent surge of using various machine learning techniques for predicting the engineering performance of concrete, most of the efforts focus on predicting single properties. However, the proper design and use of concrete in real construction require additional considerations to many other properties, such as fluidity, air-content, and constructability. In this study, we trained a deep learning neural net based on a dataset of industrial concrete, which consists of more than 10,000 samples from the production. In particular, we adopt the cutting-edge machine learning techniques to train the model to predict multiple properties of a given concrete mix design. Importantly, the results suggest that our multi-target model exhibits a higher holistic accuracy as compared to its single-target oriented counterpart. In this sense, the multi-target machine learning prediction has a strong potential promote the multi-dimensional performance optimization of concrete mix design based on the actual needs of the construction.
Semantic Segmentation of Plasma Transferred Arc Additively Manufactured NiBSi-WC Optical Microscopy Images Using a Convolutional Neural Network: Dylan Rose1; Justin Forth2; Tonya Wolfe3; Ahmed Qureshi1; Hani Henein1; 1University of Alberta; 2Consultant; 3Red Deer College
NiBSi-WC metal matrix composites (MMCs) are commonly used as an overlay material to improve the service life of components that are subject to aggressive wear environments. Plasma transferred arc-additive manufacturing (PTA-AM) offers the ability to build parts using composite materials (NiBSi-WC) to enhance service life. The wear resistance is correlated to the inherent properties of the reinforcement particles and their distribution within the metal matrix. However, the analysis of optical images to determine the weight fraction, size, and mean free path of the WC particles requires a combination of image processing and manual labeling, resulting in a time consuming and tedious task. State of the art in convolutional neural networks (CNNs) can automate this process, allowing for the distribution of carbides observed in optical microscopy images to be generated automatically. In this work, the semantic segmentation of NiBSi-WC images using a CNN architecture will be discussed.
Machine Learning in 2D Materials: Benchmarking Crystal Graph Based Convolutional Neural Network (CGCNN) for Open Databases: Shreeja Das1; Raj Kishore1; Mihir Sahoo1; S Swayamjyoti1; Anthony Yoshimura2; Nikhil Koratkar3; Saroj Nayak1; Kisor Sahu1; 1Indian Institute of Technology Bhubaneswar; 2Livermore National Laboratory; 3Rensselaer Polytechnic Institute
Identifying the right 2D materials for a targeted application is a non-trivial task because of a huge number of combinatorial possibilities. Unlike 3D bulk materials, the unavailability of large databases for 2D materials poses a unique challenge for most machine learning protocols. We employ the recently developed crystal graph convolutional neural network to benchmark some of the open databases of 2D materials for predicting both theoretical and experimental properties of 2D materials. The results indicate, bulk materials trained models are non-transferable for predicting formation energies of 2D materials. Even with a much smaller training size, 2D materials data trained models were able to capture the local chemical environment and energetics of different configurations of metal doped 2D MoS2. We also benchmark the databases in their ability to help in predicting experimental bandgaps of 18 different 2D materials. Models trained on PBE bandgaps severely underpredict optical bandgaps of materials.
Predicting Glass Behaviour from Optical Microscopy Images Using Interpretable Machine Learning: Ankur Agrawal1; Mohd Zaki1; Ravinder Bhattoo1; N. M. Anoop Krishnan1; 1Indian Institute of Technology Delhi
The deformation behavior of glasses can be classified as “normal” and “anomalous” based on the fracture pattern resulting upon indentation. When the indentation response is shear flow controlled, glasses are said to exhibit “normal” behavior. In contrast, when ring-cone cracks appear upon indentation, the glasses are said to be “anomalous”. To predict properties from microscopy images using machine learning models, we should first be able to extract features from the images. In this work, we demonstrate the ability of neural models pretrained on large datasets to extract features from microscopy images and classify the nature of glasses. Further, we also visualize those features using model explanation methods of GradientSHAP, Integrated Gradients and Occlusion. Overall, this study will guide the researchers in harnessing the capabilities of transfer learning and feature visualization for glass science.